R devel - Nov 2006 - ccf documentation bug or suggeston (PR#9394)

On 11/28/2006 11:50 AM, A.I. McLeod wrote:
> Hi Duncan,
Hi Ian.
> 
> ccf(x,y) does not explain whether c(k)=cov(x(t),x(t+k)) or
d(k)=cov(x(t),x(t-k)) is calculated.  The following example demonstrates
> that the c(k) definition is used:
> ccf(c(-1,1,rep(0,8)),c(1,rep(0,9)))
> However S-Plus acf uses the d(k) definition in their acf function.
I don't think our code looks right, let alone our documentation.  In 
both R-devel and R-patched I get a lag zero value of 1 from ccf(), and I 
think I should get cor(x,y), which in your example, would be -0.74:  and 
that is what I get from 2.4.0.

I hope this isn't a result of my patch to allow length 1 series...

Duncan Murdoch

P.S.  No need to send to more than one of R-devel, R-bugs, or me.  I'll 
see the message on any of them.  It's probably best to default to 
R-devel for suggestions like this, but the lag 0 value really is a bug, 
I think.
> 
> For interpretive purposes this is a **vital distinction** (the
cross-covariance/correlation is not symmetric like the
> autocovariance/autocorrelation).  There is not fixed convention is
textbooks or research papers.  Some define it one way and other
> another.
> 
> There is no ccf function in S-Plus. Instead there is only acf for the
auto/cross computation in multivariate time series.  This is
> more complicated since numerical output is 3D array.
> 
> Here is how S-Plus documents it:
> 
> VALUE:
> a list with the following components:
> acf
> a three-dimensional array containing the autocovariance or autocorrelation
function estimates. acf[i,j,k] is the covariance (or
> correlation) between the j -th series at time t and the k-th series at time
t+1-i.
> lag
> an array the same shape as acf containing the lags (as fractions of the
sampling period) at which acf is calculated. If j > k and i
>  > 1, then lag[i,j,k] is negative.
> 
> Ian McLeod
> 
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel

On 11/28/2006 12:50 PM, murdoch at stats.uwo.ca wrote:
> On 11/28/2006 11:50 AM, A.I. McLeod wrote:
>> Hi Duncan,
> 
> Hi Ian.
> 
>> 
>> ccf(x,y) does not explain whether c(k)=cov(x(t),x(t+k)) or
d(k)=cov(x(t),x(t-k)) is calculated.  The following example demonstrates
>> that the c(k) definition is used:
>> ccf(c(-1,1,rep(0,8)),c(1,rep(0,9)))
>> However S-Plus acf uses the d(k) definition in their acf function.
> 
> I don't think our code looks right, let alone our documentation.  In 
> both R-devel and R-patched I get a lag zero value of 1 from ccf(), and I 
> think I should get cor(x,y), which in your example, would be -0.74:  and 
> that is what I get from 2.4.0.
> 
> I hope this isn't a result of my patch to allow length 1 series...
It was.  I'll commit a patch to my patch soon, and clarify the docs at 
the same time.

Duncan Murdoch